Abstract: A speaker recognition algorithm based on vector quantization is presented.In order to obtain better recognition rate,some different characteristic parameters are adapt that can be mixed into a specific vector.In the process of VQ,traditional K-means algorithm owns the advantage of fast convergence,but it is difficult to get the global optimal result.To solve this problem,annealing algorithm is applied to improving the clusters’ quality.The details of the algorithm and related experiments are given.The results demonstrate this approach is effective.